We’re again reaching an absurd, new “uncanny valley” moment in history where the AI tools have proliferated to the masses. As a typical user, you’re confronted with obvious signs of AI use and mixed levels of effort that go into producing code and docs. That makes it harder for the humans involved to actually understand what they’ve built. In the past few weeks, I’ve been running into this dynamic more and more often, and it left me scratching my head about where it leaves us.

Picture a joint research project in which several high-profile researchers get together to define an experimental plan for validating some hypotheses. Nothing crazy about that — a typical academic project at first glance. However, it also means having to write a lot of code. Normally, that would have been the job of one or more people who are somewhat good at writing code (hard to find at universities!), with the oversight of researchers at postdoctoral level. Now, though, even the senior researchers (who’d maybe have written more code during their studies and are now focusing more on grant proposals, teaching, and academic papers) can prompt advanced LLMs to implement code for them. From hypothesis to prompt to working POC, it could literally take hours, not weeks (or even months).

This rapidly accelerated research and development cycle presents a challenge in joint understanding and decision-making.

Previously (and that means… like, last year!) you’d have half a dozen meetings over the course of several months, with researchers discussing and reaching conclusions, defining interfaces and test scenarios, someone implementing changes, validating them, and presenting the (intermediate) findings. Today, a frontier model can solve most if not all problems for you as it goes; you just provide the high-level input and make the architectural decisions. The clanker spits out code and documentation in no time. Meanwhile, you lose the ability to understand the details of the project.

The biggest absurdity is that, as a consequence, we frequently need to resort to AI to explain things. “Hey, how did you handle the case of X?” — “Huh, I don’t know, let me ask Claude” — pastes unedited Claude answer — someone has to read all of that, and it’s excruciating. I’m not even talking about this one project: it happens everywhere. I’m guilty of this too, I’m sorry, it won’t happen anymore. We’ve become meat proxies in no time, and it doesn’t feel good.

To be clear, I am not blaming people for using AI to have things implemented for them, or explained to them. The issue is when you’re no longer capable of expressing anything in your own words. That’s a problem for both you and the recipient. Have the AI explain it to you. Understand it. Then write it in your own words.

So as you keep implementing with your AI friend, new details are uncovered: unanswered questions, bugs, open decisions to be made, etc. New docs are written, yet nobody’s read them. READMEs and PR descriptions explode, with agents leaving breadcrumbs for their future selves that make sense for them, but not us. Who will decide on what to do? Who can explain the results if you’re now faced with an almost “black box” system? It’s become clear to me that we need a different way to address such projects if we, as humans, ever want to still be part of the loop, and also mentally capable of making the right decisions.

As a stopgap, I’ve resorted to two skills that help me get back on track when things are getting out of hand. They’re inspired by Matt Pocock’s skills (whose idea I like, but I won’t adopt them as-is). They’re /wait-what and /grilling. “Wait what” is remarkably simple. It lets the agent take a few steps back and explain what we’ve done, where things stand, and what the open questions are, in plain English.

---
name: wait-what
description: Re-explain the last message in plain English
---

I didn't get what you were saying, maybe I am missing context as I am jumping between lots of sessions.

Re-explain everything to me using plain ASD-STE100 Simplified Technical English, include as much context needed to understand the core issue, any gotchas/caveats, and suggested next steps.

With an almost comical form of simplicity, I find it extremely helpful during a day of context switching to have things laid out to me. Curiously, the projects I use it for feel like they’ve been running for months, but it’s only been days or weeks — it’s just their complexity paired with parallel work that necessitates frequent refreshers.

“Grilling” unravels an entire project plan and lets you decide aspects down the hierarchy, from key design decisions to minutiae. By actively thinking about the design space and ramifications of your choices, I feel like a part of what makes a good software engineer is being reactivated as you are faced with design decision questions. Because you definitely need to keep thinking critically; complacency with respect to LLM output is what’s going to fry your brain eventually, I’m sure of that.

In the longer term, I am not sure what a good setup for larger, collaborative projects looks like. Being open about using LLMs is at least a prerequisite — there’s no point pretending anymore when everyone is relying on them to keep up. The harder problem is fragmentation: everyone is using different model providers with different context. We rarely see each other’s prompts, and we don’t know which assumptions went into someone’s approach.

One solution could be a joint LLM orchestrator — perhaps a “project manager” bot — using a frontier model that everyone on the project works through, backed by the same shared knowledge: an LLM wiki paired with the code base. To make this work, people should not sidestep the agent, and everyone’s decisions and context need to pass through it. At our company, we’ve been experimenting with this approach — a marketing/sales advisor bot that we force ourselves to chat with, so everyone’s on the same page. It has been quite effective in being transparent about AI use, but it comes with new social challenges (that’s for another post).

Of course, a centralized bot only helps if it’s designed to surface trade-offs and force our joint human alignment, not again another black box. We must find effective ways to keep humans and LLMs working together, transparently, and in a way that keeps us asking less for “what did we just build?”.